Predictive residential HVAC controls can reduce a building's energy consumption; however, they require customized thermal models for each home. In this setting, detailed physical models are not practical. Fortunately, the recent availability of fine-grained thermostat data from residential buildings combined with modern machine learning creates an unprecedented opportunity to build customized data-driven thermal models. We trained and evaluated a range of promising candidate data-driven thermal models for multi-hour predictions using a sliding training window over logged temperature and equipment runtime data from 1000 smart thermostats. The models included machine learning methods, time series models, grey box models, and a simple baseline. Since many models can incorporate exogenous data, we also investigate which combination of features and history provides the best predictions of indoor air temperature. We conclude that lasso and ridge regression with solar, fan, heating and cooling runtime, and 20-minutes of history provided the lowest errors across our sample.

预测性的住宅HVAC控制可以减少建筑物的能耗；但是，他们需要为每个房屋定制热模型。在这种情况下，详细的物理模型不切实际。幸运的是，最近从住宅获得的细粒度恒温器数据与现代机器学习相结合，为构建定制数据驱动的热模型提供了前所未有的机会。我们使用滑动训练窗口对记录的温度和来自1000个智能恒温器的设备运行时间数据进行了培训，并评估了一系列有希望的候选数据驱动的热模型，以进行多小时的预测。这些模型包括机器学习方法，时间序列模型，灰箱模型和简单基线。由于许多模型可以合并外源数据，我们还研究了哪些功能和历史组合可以提供最佳的室内空气温度预测。我们得出的结论是，套索和岭回归与太阳能，风扇，加热和冷却运行时间以及20分钟的历史记录一起提供的样本误差最小。